Data is generated over many years in specialist departments, projects and systems. Each system is useful in its own right, and each data model is optimised locally. What is missing is a common context:

  • What exactly is an asset? And is it the same everywhere?
  • When are two objects the same, similar or just coincidentally named the same?
  • How are technical, organisational and specialist perspectives related?
  • Which rules apply implicitly, and where are they written down?
    •

    The answers to these questions determine whether analyses are reproducible, whether automation is possible, and whether AI systems can work reliably. This is where semantics comes into play. Semantics does not mean ‘just another data model’. It makes meanings explicit and describes terms, relationships, rules, contexts and their validity.

    In practice, we repeatedly see that the semantic deficit grows with the size and regulation of an organisation. Accordingly, the follow-up costs also increase in the form of manual coordination, special logic, exceptions and uncertainty.

    Why classic data architectures are no longer sufficient

    Many companies have responded to this situation by establishing data warehouses, data lakes, data meshes, catalogues, or metadata management. These are important steps. However, they only solve part of the problem. This is because classic architectures primarily answer the following questions:

  • Where is the data located?
  • Who is allowed to use it?
  • How is it structured?
    •

    What they do not adequately answer:

  • What does it mean?
  • How are they related in terms of subject matter?
  • What rules apply implicitly – and are they consistent?
    •

    This becomes a limiting factor, especially for advanced analytics, automation and AI. A modern LLM or AI agent may be very good at reading texts, making plans and using tools. However, without explicit semantics, reliability is lacking. In other words, without a semantic layer, AI is impressive but not resilient.

    The semantic layer as the missing layer

    We are convinced that the semantic layer forms a central layer of modern data and AI architectures. Not as a monolithic ‘master model’, but as a living semantic system. This layer consists of:

    Ontologies, taxonomies, vocabularies;

  • Knowledge graphs;
  • Rules and conditions;
  • References to standards and norms;
  • Links between structured and unstructured data.

    • The crucial point here is that the semantic layer combines human knowledge with machine usability. It is therefore the bridge between specialist areas and IT, documents and databases, rules and exceptions, as well as the past (established systems) and the future (automation).

    But this is precisely where many organisations reach their limits.

    This is because semantics are often complex to model, difficult to keep consistent, highly domain-dependent and often poorly documented historically.

    This is where Agentic AI comes into play.

    What does Agentic AI have to do with semantics?

    Agentic AI is often described in terms of autonomy, planning and tool usage.

    But at Alexander Thamm [at], we see the real added value of Agentic AI in something else: it can scale semantic work.

    Specifically, this means:

  • Agents can analyse large document collections and extract terms.
  • They can compare data structures, recognise patterns and flag inconsistencies.
  • They can generate suggestions for classes, relations and mappings.
  • They can check existing data models against rules and find deviations.
  • They can simulate the effects of changes in semantics.
    •

    It is important to note that AI agents do not replace expert decisions. Instead, they take on the work that humans are poor at scaling, i.e. mass analysis, time-consuming preliminary checks and tedious consistency checks.

    This shifts the focus for humans away from manual drudgery and towards expert evaluation, governance and quality.

    The semantic layer of the future is agentic

    From our point of view, a new architectural principle is emerging here: the semantic layer of the future will be an agent mesh. The reason is obvious, because semantics is not a static construct. Terms change, standards evolve, organisations restructure, and systems are added as quickly as they disappear. A static knowledge graph cannot reflect this dynamic.

    Instead, it requires the interaction of specialised agents: agents that monitor models and classify new data, agents that check rules and reveal inconsistencies, and agents that make comprehensible and verifiable suggestions. Only such a living, agent-based system can keep semantics in organisations permanently up to date, consistent and usable.

    This makes the semantic layer active instead of passive, checking instead of merely descriptive, and evolutionary instead of modelled once and for all. That is why agentic AI and semantics are inseparable.

    Governance first – otherwise it won't scale

    One aspect is central and often underestimated: governance. Agents working on semantics delve deep into a company's knowledge base – and without clear guidelines, this can quickly become risky. Our experience therefore shows time and again that governance must come before autonomy. Roles, approvals and quality barriers must be clearly defined, and a human-in-the-loop is not an optional convenience feature, but a requirement. Likewise, decisions must be explainable and auditable at all times.

    Only under these conditions can genuine trust be established, both internally and externally.

    Conclusion

    Many companies are faced with the question of how they can integrate AI into their organisation in a productive, secure and scalable manner. Our answer to this is clear: without semantics, there can be no robust AI, and without agents, there can be no scalable semantics. Agentic AI and semantic layers are therefore not separate developments, but two sides of the same coin.

    Combining the two creates robust automation, traceable decisions and a knowledge base that grows alongside the organisation.

    That is exactly what we are working on.

    Sources

    1) Sequeda et al., Knowledge Graphs as a source of trust for LLM-powered systems (2025) https://www.sciencedirect.com/science/article/pii/S1570826824000441

    2) Jaber et al., AutoClimDS: Climate Data AI (arXiv 2025) https://arxiv.org/abs/2509.21553

    3) Peshevski et al., AI Agent-Driven KG Construction (arXiv 2025) https://www.arxiv.org/abs/2511.11017

    4) McGee et al., Enabling Ethical AI with Ontological Context (arXiv 2025) https://arxiv.org/abs/2512.04822

    5) Open Research Knowledge Graph: The Open Research Knowledge Graph (ORKG) aims to describe research papers in a structured manner. With the ORKG, papers are easier to find and compare. https://orkg.org/

    6) Semantic layer – Wikipedia (Def.)
    https://en.wikipedia.org/wiki/Semantic_layer?utm_source=chatgpt.com

    About the author

    Dr Andreas Kyek is a data science and AI expert with over 25 years of experience in data-driven product and process development. With his background in physics and his work in leadership roles (including at Infineon), he combines technological depth with strategic implementation. As Senior Principal Data Scientist and Practice Lead at Alexander Thamm [at], he is expanding the data science and AI practice, focusing on agentic AI systems, multi-agent architectures, semantic knowledge models and RAG in complex industrial setups. He leads large-scale data/AI initiatives (industry, energy, mobility, infrastructure) and is involved in mentoring and training for the responsible use of AI.